>> GO MORNING AND WELCOME TO THE NEXT OF THE BIOWULF SEMINAR SERIES. WE'RE HAPPY TO PRESENT DR. CHUANG WHO RECEIVED HIS DEGREE FROM BOSTON IN 2009 AFTER THAT HE WENT TO DO A POST-DOC AT THE UNIVERSITY OF STRASBOURG BEFORE COMING TO NIH. IN 2011 HE CAME TO NIH AND JOINED THE STRUCTURE IN PETER QUAN'S GROUP. HE BECAME THE CO-HEAD OF THE SECTION. HIS CURRENT RESEARCH FOCUSES ON THE DEVELOPMENT AND APPLICATION OF COMPUTATIONAL TOOLS FOR ANTIBODIES TARGETING HIV 1 AND OTHER VARIOUS PATHOGENS. >> THANK YOU. GOOD MORNING. FIRST I WOULD LIKE TO THANK SUSAN AND THE BIOWULF TEAM. I'M A STAFF SCIENTISTS AND CO-HEAD OF THE BIOFORMATTICS AND TALK ABOUT ANTIBODY RESISTANT DETECTION. FIRST I WANT TO GIVE YOU A GENERAL PICK TIRE -- PICTURE OF OUR CENTER. IT WAS SET UP 18 YEARS AGO AIMING TO DEVELOP THE HIV VACCINE BUT DURING THE TIME WE HAVE BEEN EXPANDING OUR PROGRAM TO DEVELOP VACCINE TARGETING FOR INFLUENZA AND MALARIA AND OTHERS AND IN ADDITION TO DEVELOPING VACCINES WE'RE INTERESTED IN BROADENING ANTIBODIES THAT CAN SERVE AS THERAPIES FOR HIV 1 AND OTHER PATHOGENS. WE DO RESEARCH AND ONCE WE DEVELOP AN INTERESTING VACCINE OR ANTIBODY CANDIDATE WE MOVE TO THE PRODUCTION PLANT AND WE HAVE A GMP MANUFACTURING FACILITY AT FREDERICK WHICH PRODUCES GMP QUALITY MATERIALS FOR CLEAN CULTURELLS AND AFTER WE EVALUATE FOR EXAMPLE NEUTRALIZATION RESPONSE AT A VACCINE TESTING LABORATORY IMMUNO LOGIC TESTING LABORATORY AND THE BASIC RESEARCH FOR OPTIMIZING THE CONSTRUCT. SO IN TERMS OF THE STRUCTURAL BIOFORMATICES IT TO PERFORM STRUCTURE ANALYSIS AND PROTEIN DESIGN TO ASSIST THE DEVELOPMENT OF VACCINES IN THERAPEUTIC ANTIBODIES. TODAY I'M GOING GIVE YOU TWO PLAYERS OF OUR RESEARCH RELATED TO BOTH VACCINE AND ANTIBODY. FIRST I'M GOING TO TALK ABOUT IMMUNOSTABILIZATION AND SHOW YOU A COUPLE EXAMPLES WHERE WE APPLY COMMUTATIONAL METHODS TO STABILIZE THE HIV TRYMER AND STABILIZING PROTEIN CONFIRMATION B DESIGN AND THE SECOND HALF I'LL TALK ABOUT WHAT WAS DEVELOPED FOR PREDICTING ANTIBODY RESISTANCE FOR THE HIV 1 BROADLY NEUTRALIZING ANTIBODIES. SO FIRST IT'S THE IMMUNEGENAL STABILIZATION. THERE'S MANY THAT WERE WELL DEVELOPED BEFORE THE KNOWLEDGE OF THEIR ANTIGEN STRUCTURE. AND WE REALIZED TO SEE PATHOGENS SUCH AS HIV 1 AND OTHER VIRUSES AND WE NEED TO HAVE STABILIZE THE AND FOR HIV 1 AND IT FACES FLEXIBILITY THROUGH THE TRYMER ON THE VIRUS WHICH IS THE ANTIGEN ON THE VIRUS SURFACE WHEN IT'S IN A LIGAND FREE STATE IT HAS A CONFIRMATION BUT WHEN THE TRIMER BINDS IT RECEIVED CD4 BOUND CONFIRMATION AND WHEN IT TRAPS THE CELLS IT RECEIVES AUTOCONFIRMATIONS SO THE TRIMER MOVES BETWEEN CONFIRMATION. RECENTLY WE HAVE SHOWN FOR THE BROADLY NEUTRALIZING ANTIBODIES ABLE TO NEUTRALIZE THE VIRUS CLOSE CONFIRMATION WHILE THE WEAK OR NON-NEUTRALIZING ANTIBODY ARE NOT COMPATIBLE WITH THIS FORMATION BUT COMPATIBLE WITH OTHER CONFIRMATION. IT SUGGESTS IF YOU WANT TO DEVELOP A VACCINE THAT TARGETS -- THAT IS ABLE TO NEUTRALIZE THE ANTIBODY BASED ON THE TRIMER YOU WANT TO STABILIZE IT IN THE CLOSED STATE. I'M GOING TO USE ANOTHER VIRUS AS AN EXAMPLE WHICH IS RSV WHICH CAUSES RESPIRATORY SYNDROMES AND IS A MAJOR CAUSE OF INFANT HOSPITALIZATION IN THE UNITED STATES. THE FUSION PROTEIN OF THE VIRUS HAS TWO CONFIRMATIONS. THE PREFUSION CONFIRMATION AND POST-FUSION CONFIRMATION. SO RESEARCHERS HAVE SHOWN THE NEUTRALIZING ACTIVITY OF TARGETING THIS VIRUS HAS PREDOM NANLT -- PREDOMINANTLY PRE-FUSION STATUS AND YOU WANT TO STABILIZE IT IN A PROTEIN. THIS IS NOT A TRIX -- TRIVIAL THING TO DO BECAUSE IN ONE LAB WE WERE ABLE TO OVERCOME THE PROBLEM BY STABILIZE THE STRUCTURE INFUSION CONFIRMATION STABILIZED BY AN ANTIBODY. AND IT CAN STABILIZE IN A SPECIFIC CONFIRMATION. BADE -- BASED ON THIS STRUCTURE WE WERE ABLE TO FILL CAPACITY FILLING MUTATIONS TO STABILIZE IT IN A PRE-FUSION CONFIRMATION. WE HAVE SHOWN IN THE STUDY USING AN ANIMAL MODEL WE WERE ABLE TO SHOW THE CONSTRUCTING IN THE PRE-FUSION SUCH AS THE DS OR DS CAP WAS HERE IT WAS TIGHTER COMPARED TO THE POST-FUSION FORM OF THE IMMUNOGEN. THIS WORK WAS RECOGNIZED AS A RUNNER-UP FOR THE BREAKTHROUGH OF YEAR IN 2013 AND SUGGESTING IF YOU WANT TO DEVELOP AN IMMUNOGEN YOU MAY WANT TO ELICIT THE HIGHEST NEUTRALIZATION RESPONSE. HOPEFULLY I HAVE CONVINCED YOU IN CERTAIN CASES IMMUNOGENIC STABILIZATION CONFORMATION IS USED A STRUCTURE-BASED APPROACH AND I'LL TALK TO YOU ABOUT IT FOR A COUPLE OF THESE. SO THE SERVICE AND MOST OBVIOUS THING PEOPLE DO IS SOFTWARE DESIGN. AND IF YOU DESIGN THE PROXIMAL AND OPTIMAL GEOMETRY YOU CAN FORM THE STABILIZER FOR THE NEARBY DOMAINS. IN TERMS OF THE SUCCESSFUL DESIGN IN ADDITION TO THE IMMUNOGEN THERE'S ANOTHER TRIMER WHICH WE ENGINEERED A DISULFIDE FOR CLOSE CONFIRMATION AND THERE'S A NICE WEB SERVER AND THIS IS WHERE YOU CAN PUT IN THE PAIRS AND THIS LOOKS SIMPLE SO WHY DO WE CARE ABOUT THE OTHER MASSES? THERE'S A COUPLE DRAWBACKS FOR THE DISULFIDE DESIGN. LOOK AT THE STRUCTURE AND SAY YOU CAN SEE THE TWO CYSTEINE AND THEY'LL CAN CAUSE THE PROTEIN TO FOLD INCORRECTLY AND THE SECOND DRAWBACK FOR DISULFIDE DESIGN IS BECAUSE DISULFIDE IS A COVALENT BIND. WHEN WE TRY TO STABILIZE THE MUTATIONS WE HAVE TO CONSIDER OTHER MASSES IN ADDITION TO THE SIMPLE DISULFIDE DESIGN AND I'LL LOOK AT THE TRIMER AND THIS CAN BE ACHIEVED BY OPTIMIZING THE INTERFACE OF FLEXIBLE DOMAINS. I'M SHOWING YOU THE HIV1 TRIMER AND THERE'S A TARGET CONFIRMATION. YOU CAN SEE THE LOOPS ARE SORT OF STACK NICELY TOGETHER. WELL, FOR THIS ALTERNATIVE CONFIRMATION SUCH AS CD4 CONFIRMATION THEY'RE PART FROM EACH OTHER SUGGESTING IF YOU WANT TO STABILIZE THINGS IN A CLOSE CONFIRMATION ONE WAY TO DO THIS IS TO LOOK AT THE TRIGGER THE CONVERSION TO THE OTHER STATE. SO THIS IS A COMPUTATIONALLY INTENSIVE PROGRAM BECAUSE TYPICALLY WE LOOK AT RESIDUE POSITIONS. EACH RESIDUE THERE IS THE COMPLEXITY OF DIFFERENT AMINO ACID TYPE AND ALL IN ALL IT'S A COMPUTATIONAL INTENSIVE PROBLEM. EVEN THE EXPERTS IN THIS CASE CANNOT DO THIS MANUALLY. YOU HAVE TO RELY ON A COMPUTATIONAL ALGORITHM. AND DUE TO THE COMPUTATIONAL COMPLEXITY IN THIS CASE WE DON'T USE ENERGETIC CALCULATIONS INSTEAD WE TYPICALLY USE PROTEIN DESIGNS WITH A UTILIZE FUNCTIONS SUCH AS A ROTAMER. WE HAVE LOOKED AT THE V1, V 2 AND V3 AND YOU CAN SEE THE RESIDUES WHICH ARE ENNER GETCALLY NOT FAVORABLE. WE USE PROTEIN DESIGN WITH THE ABILITY TO SAMPLE ALL ROTAMER COMBINATIONS AND IF IT'S EFFICIENT AND WE USED AMINO ACIDS COMBINATION AND YOU CAN SEE THERE'S AN OPTIMIZE IN INTERACTION WITH A PATCH WITH IMPROVED CONTEXT. SO OFF THE WE DID THE FOUR INTERFACE MUTATIONS WE WANTED TO EVALUATE HOW IT IMPACTS THE IMMUNOGENETICY SO IN THIS STUDY WE EVALUATED THREE DIFFERENT HIV TRIMER CONSTRUCT. FOR THE SAKE OF THE TALK CONSIDER IT LIKE A WILD TIME OR BASELINE TRIMER. THE SECOND IMMUNOGEN IS THE DISULFIDE AND THE THIRD CONSTRUCT IS THE ADDITIONAL MUTATIONS ON TOP OF THE DS-SOSI PIVE AND WE LOOKED AT ANTIBODIES AND NON-HIV OR NON-NEUTRALIZING ANTIBODY. AND WE USED THE MSD TO LOOK AT THE TRIMER AND IN ALL THREE CASES THEY HAD SIMILAR CHARACTERISTICS SO THERE'S NO DIFFERENCES HERE BUT IF YOU LOOK AT BINDING TOWARDS THE ANTIBODY WHICH WE WANT TO REDUCE, THE DS HAS LESSER IMMUNOGENICITY AND SUGGESTING WITH THE ADDITIONAL FOUR MUTATIONS TO STABILIZE THE TRIMER IN THE PRE-FUSION TARGET STATE AND WE EVALUATED THE CONSTRUCT USING THE SCANNING AND YOU CAN SEE WITH THE COMBINATION IT INCREASES COMPARED TO THE WILD TYPE TRIMER. SUGGESTING THE OVERALL STABILITY OF THE TRIMER WITH THE FOUR MUTATION WAS INHANSED. -- ENHANCED. AND WE GAVE THE FOUR CONSTRUCT TO GUINEA PIG AND THE VARIANT WITH THE MUTATIONS ELICITED ABOUT A 20-FOLD INCREAS OF THE MUTATION TRIMER COMPARED TO THE WILD TYPE TRIMER SUGGESTING BY STABILIZING THE TRIMER THE CONFIRMATION YOU'RE TARGETING YOU'RE ABLE TO ENHANCE THE IMMUNOGENICITY AS WELL. SO THIS IS WHAT WE HAD DONE FOR THE STABILIZATION OF HIV TRIMER NOW I'M GOING TO TALK ABOUT AN AUTOMATED TOOL FOR DESIGN. SO WHAT IS SPECIAL ABOUT PROLLIN COMPARED TO OTHER AMINO ACIDS IT ASSUMES A DIFFERENT CONFIGURATION AND THE OTHER AMINO ACIDS HAVE TWO IMPLICATION. FIRST, PROTEIN HAS FLEXIBILITY SO IN GENERAL IF YOU IT TO THE PROTEIN IT WILL MAKE PROTEIN MORE STABLE UNLESS YOU PLACE IT IN A POSITION THAT IS NOT COMPATIBLE AND DUE TO THE SAME REASON IT ASSUMES A DIFFERENT AMINO CONFIGURATION, IF IT'S NOT AVAILABLE FOR THE OTHER AMINO ACIDS. SO YOU CAN PLACE A PROLINE IN A REGION ONLY PRESENT IN THE ALTERNATIVE CORM FORMATION TO BREAK OUT THE HELIX TO SHIFT THE BALANCE AND THE CONFORMATION YOU ARE TARGETING. SO PROLINE HAS BEEN SUCCESSFUL pYO -- GLYCOPROTEINS AND IN GLYC PARALLEL TO OTHER PREFUSION IMMUNOGEN A GROUP AND THE FUSION PROTEIN WAS IN THE PRE-FUSION CONFIRMATION AND GIVES INCREASED EXPRESSION AND THERE ARE MANY EXAMPLES SUCH AS WHICH HAVE BEEN STABILIZING IN THE PRE-FUSION FORM USING PROLINE AND THE VIRUS FUSION PROTEIN, HIV-1 TRIMER AND THE PROTEINS AND INTERESTING THERE ARE NO PUBLICLY AVAILABLE TO US THAT CAN PERFORM THESE TYPES OF PROTEIN DESIGNS. SO IT'S WHY WE DEVELOPED THE AUTOMATED PIPELINE FOR PROTEIN CONFORMATION BY PROLINE. SO THE PROGRAM DETERMINES IF A RESIDUE POSITION IS COMPATIBLE WITH THE PROLINE IN THE FOLLOWING WAY. SO IF A RESIDUE POSITION IS IN THIS POSITION WE SAY THIS POSITION IS NOT COMPATIBLE WITH THE PROLINE SO IF THE ANGLES OF THE RESIDUE POSITION IS NOT COMPATIBLE WITH THE PROLINE ANGLES THEN THIS POSITION IS NOT COMPATIBLE WITH THE PROLINE. FINALLY, WHEN YOU INTRODUCE THE PROLINE IT CAUSES CLASHES WITH THE OTHER RESIDUE POSITIONS AND THEN THIS POSITION IS ALSO NOT COMPATIBLE WITH THE PROLINE. IT'S ONLY WHEN THE POSITION IS NOT A HELICAL CONFORMATION TO ALLOW BACKBONE ANGLES AND NOT CLASHES WITH OTHER RESIDUES AND CONSIDER IT POTENTIALLY COMPATIBLE WITH THE PROLINE. SO THE PROGRAM CAN POTENTIALLY TAKE THE TARGET CONFIRMATION OF INPUT OR CONFIRMATION'S INPUT OR BOTH. SO IF ONLY HAVE A TARGET CONFORMATION THE PRE-FUSION FORM OF THE STRUCTURE THE PROGRAM GIVES A LIST OF RESIDUES COMPATIBLE FOR THE PROLINE TO DESIGN IT TO STABILIZE IT IN THE CONFORMATION. IF HAVE YOU AN ALTERNATIVE CONFORMATION THE POST-FUSION FORM OF THE PROTEIN, THE PROGRAM SPITS OUT THE RESIDUE IS NOT COMPATIBLE TO STABILIZE IN THE TARGET CONFORMATION. AND FINALLY, AND THE BEST CASE SCENARIO IS YOU HAVE INFORMATION FOR BOTH THE TARGET CONFORMATION AND IN THIS CASE THE PROGRAM DESIGN A PROLINE.COMPATIBLE TO - AND THIS RESIDUE HAS THE HIGHEST AMOUNT OF SUCCEEDING TO STABILIZE THINGS IN YOUR TARGET CONFORMATION. SO AFTER THE IMPLEMENTATION OF THE PROGRAM WE FIRST USED THE PRE FUSION FORM OF THE RSV APPLICANT TO SEE IF WE WERE ABLE TO IDENTIFY THE RESIDUES USED TO STABILIZE THE PRE-FUSION CONFORMATION AND WE SUCCESSFULLY DID THAT AND FROM THE STRUCTURE YOU CAN SEE WHY THIS PROLINE MUTATION WORKS BECAUSE IN THE POST-FUSION FORM THIS RESIDUE RESIDES IN THE PELICO REGION AND IT SUGGESTS IF YOU PUT THE PROTEIN HERE IT WILL BREAK OUT IN POST-FUSION CONFORMATION BUT WON'T EFFECT THE GEOMETRY OF THE PRE-FUSION FORM. THIS IS WHY THE MUTATION WAS ABLE TO DO ITS WORK. SO WE ALSO APPLIED THIS TO OTHER KNOWN SYSTEMS WHICH THERE ARE PROTEIN MUTATION THAT TARGETS THE MUTATION AND ARE ABLE TO IDENTIFY THE RESIDUES AS ONE OF OUR HITS. WE UPLOADED OUR PROGRAM SO YOU ARE FREE TO DOWNLOAD IT AND TEST IT YOURSELF AND PERFORM SOME TYPE OF PROTEIN STABILIZATION OF YOUR OWN. TO SUM THIS UP PRETTY QUICKLY, FIRST I INTRODUCED TO YOU THE ANTIGENIC FLEXIBILITY AND OFTEN IT IS GOOD TO STABILIZE IT IN A PARTICULAR CONFORMATION. I SHOWED A COUPLE EXAMPLES WE CAN PERFORM PROTEIN CONFORMATION STABILIZATION SPECIFICALLY WE WERE ABLE TO IDENTIFY MUTATION AND ABLE TO STABILIZE THE HIV-1 TRIMER AND WE SHOWED WITH THESE ADDITIONAL THERMAL STABILITY AND THE IMMUNOGENICITY OF THE DESIGN AND ALSO SHOWED YOU THE NEW PROTOCOL WE DEVELOPED TO AUTOMATICALLY DESIGN A STABILIZING PROTEIN CONFIRMATION. WE'RE HAPPY TO SET UP THE PROGRAM AND WE'RE MORE THAN HAPPY TO HELP YOU DESIGN PROTEIN STABILIZATION TO HELP IMPROVE THE PROPERTIES OF YOUR PROTEIN CONSTRUCTS. SO IN TERMS OF FUTURE DIRECTIONS, FIRST WE ARE APPLYING ALL THE TECHNIQUE TO STABILIZE OTHER PROTEINS OF INTEREST. FOR EXAMPLE, THE GLYCOPROTEIN FOR INFLUENZA 3 AND OTHERS AND WITH THE MUTATIONS WE IDENTIFIED WE ARE TRANSFERRING THE MUTATIONS TO IMMUNOGEN CONSTRUCTS AND WE HAVE SHOWN THIS MUTATION IMPROVES THE OTHER HIV-1 TRIMERS AS WELL. AND FINALLY, WE ARE FURTHER IMPROVING THE PROGRAM TO FURTHER PRIORITIZE OUR PREDICTIONS. SO I'M WONDERING IF THERE'S ANY QUESTIONS AT THIS MOMENT. [QUESTION OFF MIC] >> FOR THIS PROGRAM IN TERMS OF STRUCTURAL COMPATIBILITY WE ONLY LOOKED AT CLASHES. BECAUSE THIS IS SORT OF LIKE A -- WE'RE NOT SAYING ALL THE MUTATIONS WE IDENTIFY ARE GOING TO WORK OR BE ENERGETIC BUT IT'S TO A REASONABLE NUMBER OF RESIDUES TO TEST IN THE LABORATORY AND WE DECIDE TO USE A SUBFIT -- SUBFILTER TO LOOK AT THE PROLINES BUT FOR EXAMPLE, THE PROLINE HELIX WILL BREAK UP SO THERE'S NOT MUCH ISSUE BUT YOU MAY HAVE A HIT OR MISS HERE OR THERE SO FOR THAT PART WE HAVE USED A FILTER. [QUESTION OFF MIC] >> YES. SO THE LIST, YOU MEAN THE HITS FROM A PROGRAM OR VALIDATED ONES? SO THERE ARE SIX, YES, CORRECT. IT'S ONE OF THE 63. [CONVERSATION OFF MIC] >> TO TELL YOU THE TRUTH THE REASON -- WE MIGHT DO SO TO DEMONSTRATE OUR CAPABILITY OF PROGRAM IF WE HAVE THE RESOURCE BUT TO BE FRANK WITH YOU, THIS IMMUNOGEN IS ALREADY MOVING FURTHER ALONG TO PHASE 1 STUDIES. WE COULD DO MORE TO VALIDATE THE OTHERS BUT IT MAY BE TOO LATE TO CATCH UP WHAT IS BEING DEVELOPED. IT'S A GOOD SUGGESTION. WE COULD FURTHER DO THAT. [COMMENT OFF MIC] AND WE'RE MOSTLY FOCUSSED ON PRE-FUSION CONFORMATION VERSUS POST-FUSION CONFORMATION. IF YOU LOOK AT WHAT WAS MEANT BY THE ANTIBODIES WE'RE PREDOMINANTLY LOOKING AT THE REGION OR TRYING TO ELICIT ANTIBODY TARGETS FOR THE FUSION PROTEIN. I KNOW IN THE LITERATURE THERE ARE MORE AND MORE ANTIBODIES HONE -- SHOWN TO TARGET THE INTERFACE BETWEEN THE PROTEMER AND THAT IS ANOTHER WAY TO DESIGN THE IMMUNOGEN BUT I KNOW THERE'S SOME EFFORTS ON THIS. ARE THERE ANY QUESTIONS ON THIS PART? OKAY. SO LET ME CONTINUE TO THE SECOND HALF OF THE TALK WHICH IS ON PREDICTING ANTIBODY RESISTANCE. THERE'S BEEN MANY ANTIBODIES ISOLATED FROM THE SERUM OF CHRONICALLY INFECTED PATIENTS AND THEY TARGET ALL MAJOR ACCESSIBLE REGIONS ON THE HIV-1 ENVELOPE TRIMER AND THEY COULD OFFER PROMISING UTILITY FOR PREVENTION AND TREATMENT OF HIV-1 INFECTION. THE ISSUE IS HIV-1 YOU KNOW IS DIVERSE AND SUCH THAT IT CANNOT SIMPLY NEUTRALIZE ALL STRINGS AND I'M SHOWING IT ONLY NEUTRALIZED 90% AND THERE'S 10% OF CASES THIS ANTIBODY WILL JUST NOT WORK. SO IT WOULD BE GREAT IF THERE IS A TOOL TO TELL US IF THE SERG -- CIRCULATING VIRUS IS SENSITIVE TO THIS ANTIBODY SO THAT WE CAN KNOW IF WE WANT TO USE IT TO TREAT THE PATIENT OR WE SHOULD USE ANOTHER ANTIBODY. SO HIV BECAUSE IT'S AN RNA VIRUS IT CAN ACCUMULATE THE MUTATION QUICKLY DUE TO THE HIGH MUTATION RATE. EVEN IF YOU STARTED WITH THE CIRCULATING VIRUS SENSITIVE TO THE ANTIBODY AND YOU TAKE THE ANTIBODY FOR TREATMENT AND DURING THE PROGRESS SOME OF YOUR VIRUS WILL GROW RESISTANCE TO THE ANTIBODIES YOU'RE ADMINISTRATING AND YOU HAVE TO CHANGE TREATMENT OTHERWISE THE ANTIBODY WILL NOT BE EFFECTIVE. IT WOULD BE A GREAT TOOL. THERE'S A TOOL THAT CAN MONITOR THE PROGRESS OF THESE VIRUSES SUCH THAT YOU KNOW WHEN THE RESISTANCE OCCURS AND WHEN TO CHANGE THE TREATMENT. SO FOR THE ANTIBODY OR DRUG RESISTANCE THERE'S BASICALLY TWO COMMON METHODS. FIRST IS PHENOTYPICALLY LIKE THE GOLD STANDARD BUT IN THIS CASE YOU HAVE TO PRODUCE THE VIRUSES. ONCE YOU KNOW THE SEQUENCE YOU HAVE TO PRODUCE THE VIRUSES AND PRODU PRODU PRODUCE IN VITRO ASSAYS AND WHAT PEOPLE USUALLY DO FOR DRUG RESISTANCE BECAUSE WE'RE FACING EXPERIENCE FROM THE DRUG-RESISTANT FIELD SO PEOPLE HAVE BEEN DOING FOR THE HIV DRUG-RESISTANT IS GENO TYPICALLY BASED ON THE RESISTANCE GATHERED FROM EMERGING CLINICAL SAMPLES OR FROM MUTATION WERE DERIVED FROM CELL CULTURE AND OTHER PRESSURES AND THERE'S A SET OF KNOWN MUTATION ARE RESISTANT. ONE CAN LOOK AT A SEQUENCE OF THE PATIENT SAMPLE AND SEE IF THERE'S MUTATIONS AND DETERMINE IF IT'S RESISTANT OR SENSITIVE TO THE ANTIBODY BUT THE INTERPRETATION CAN BE TRICKY BECAUSE THEY DO NOT ACT INTERDEPENDENTLY. SO WHEN THE INFORMATION IS STARTING TO GROW IT'S DIFFICULT TO INTERPRET THE RESULTS BASED ON A SET OF MUTATION. THE GOAL OF THE STUDY IS TO DEVELOP MACHINE-LEARNING ALGORIT ALGORITHM USING THE VIRUS SEQUENCING INPUT. SO HERE WE USE THE LEARNING PARADIGM WHICH IS TO LEARN THE FUNCTION AND MATCH THE INPUT AND OUTPUT BASED ON THE PAIRS. THE INPUT IS THE HIV-1 ENVELOPE SEQUENCE. THIS IS THE PROTEIN ON THE VIRAL SURFACE AND FOR THE INPUT FEATURE IS ENCODING OF AMINO ASSAY TYPES. IT WAS ONE TYPE AND IF THERE'S RESIDUE IN POSITION 1 WE WOULD PUT THE 1 FOR RESIDUE 1 AND 0 FOR THE OTHER AMINO ACID TYPE. IN ADDITION TO THE STANDARD 20 AMINO ACID WE ALSO INTRODUCED A NEW AMINO ACID TYPE AND ANOTHER WHICH DESIGNATES IF THERE'S GLYCOL RESIDUE OR NOT. THE CONSEQUENCE IS MORE THAN 800 RESIDUES SO SLIGHTLY LESS THAN 2,000 FEATURES WE ARE USING TO TRAIN OUR MODEL. SO THAT'S OUR INPUT. FOR THE OUTPUT WE USE A BINARY OUTPUT TYPE WHICH IS IF THE VIRUS IS SENSITIVE OR RESISTANT TOWARDS THE SEQUENCE AND WE USE IT THE 50 MICROGRAM PER ML WHICH IS A STANDARD IN THE FIELD AND WE PUT IT ANYONE THE DATABASE HOSTED BY THE LOSSAL -- LABORATORY AND THE CHOICE IS A GRADIENT BOOSTING MACHINE WHICH CAN BE TRAINED IN A REASONABLE TIME FRAME WHILE IT OFFERS ACCURACY HIGHER THAN A RANDOM MODEL. SO IT MINIMIZING THE LOSS FUNCTION AND IF YOU WANT TO LOOK AT THE DATA CODES THE GREEN FROM THE RED. THE PROGRAM WILL FIRST CREATE A WEAK LEARN USUALLY LIKE A TREE STUMP AND WILL TRY TO SEPARATE THE GREEN FROM THE RED AND THAT COULD GENERATE MISCLASSIFIED INSTANCES. IT WILL PROBABLY TRY TO FIND ANOTHER WEAK CLASSIFIER BUT IF YOU ADD THESE TWO MODELS TOGETHER IT WILL ACTUALLY GIVE YOU A HIGHER PREDICTION ACCURACY AND CLASSIFY INSTANCES. WHILE THE METHOD IS ADD MORGUE -- ADDING MORE AND MORE CLASSIFIERS. WE CALL IT THE BROADLY NEUTRALIZING ANTIBODY PREDICTOR. WE TRAIN THE PREDICTORS FOR 33 ANTIBODIES USING A TRAINING SET THEY'RE COLOR CODED DEPENDING ON WHERE THEY RECOGNIZED IN THE ENVELOPE AND TRIMER AND I'M SHOWING THE PERFORMANCE ON THE CURVE OF THE CROSS-VALIDATION FOR THIS PREDICTOR. YOU CAN SEE THE MEDIUM AUC IS 8.9 WHICH IS QUITE NICE AND THE NICE THING ABOUT THIS IS THE PREDICTIONS ARE GENERALLY GOOD FOR ALL TYPES OF ANTIBODIES. IT DOESN'T FAVOR MUCH BERTD -- BETTER FOR CERTAIN ANTIBODIES AND WILL GENERATE IN A REASONABLE FASHION. THERE'S NO PREFERENCE. AND WE CAN USE THE ALGORITHM AND WE'RE ABLE TO OBSERVE FOR THE 28 OF THE 33 CASES OR PERFORMING BETTER THAN THE SECOND BEST PREDICTOR FROM THE THREE SUGGESTING WE'RE CHOOSING THE BEST OF THE MACHINE LEARNING ALGORITHM WE CHOSE IS GOOD AND WE'RE TESTING OTHER METHODS AND THERE MAY BE AN UPDATE ON THIS. SO AFTER WE KNOW THAT OUR TRAINING WAS FAIRLY WIDE WE WANT TO TO KNOW IF THE PROGRAM WILL WORK IN A REAL-LIFE SCENARIO. WE WANTED TO SEE HOW IT WORKS ON PREDICTING. IT PREDICTS VERY WELL ON A NUMBER OF CASES. I WANT TO SHOW YOU ONE STUDY. HERE WE EVALUATED THE ANTIBODY RESISTANCE FOR 140 CLINICAL ISO LATES WHICH IS A TRIAL TO EVALUATE THE SAFETY AND ANTIBODY VRCO1. IF YOU LOOK AT THE PREDICTION AT A SEQUENCE LEVEL, THERE IS SO OF 140 WERE CONSIDERED RESISTANT WHEN CONSIDERING IN VITRO AND LOOK AT THE STRAINS ISOLATES AND WE WERE ABLE TO CUL THE MAJORITY AS SENSITIVE. THERE'S ONLY A FEW POSITIVES THERE. SO YOU CAN SEE IN THE FIRST LEVEL THE PROGRAM PERFORMED PRETTY WELL. WHAT ABOUT THE PATIENT'S LEVEL. IF YOU ASK THE QUESTION ANOTHER WAY, THE 140 ISO LATES COME FROM ONLY EIGHT PATIENTS. ONLY TWO OF THE EIGHT PATIENTS PATIENT 21 AND 25 HAVE RESISTANCE SEQUEL OR A PROGRAM IS ABLE TO TELL YOU THAT THESE TWO PATIENTS HAVE RESISTANT SEQUENCES WHILE SIX OTHERS DID NOT DEVELOP RESISTANCE. IF WE LOOK AT THE DATA IN THE AGGREGATE VIEW BASED THROWN PATIENT LEVEL WHERE ACCURACY IS 100%. SO IN PREDICTING ANTIBODY RESISTANCE WITH HIGH ACCURACY. ONE NICE THING ABOUT GPM AND OTHER METHODS YOU CAN OBTAIN YOU CAN KNOW WHICH FEATURES ARE BASICALLY RESIDUES FOR ANTIBODY RESISTANCE. HERE I'M SHOWING YOU A TABLE OF THE TOP THREE FEATURE WITH THE HIGHEST FOR ANTIBODIES A1C95 AND THE THREE RESIDUES ARE RELATE TO GLYCAN 234 AND 276. YOU CAN SEE BASED ON THE STRUCK STRUCTURE AND THE DATA WE KNOW THAT THE TWO GLYCAN CRITICAL FOR THE BINDING AND NEUTRALIZATION OF THE ANTIBODY. SO IT IS NICE TO SEE THAT ALSO FROM THESE PREDICTORS WE'RE ABLE TO IDENTIFY THE SAME SET OF FEATURES. AND THEY'RE QUITE IMPORTANT TO HELP IS UNDERSTAND THE MECHANISM OF ANTIBODY RESISTANCE ESPECIALLY IN THE CASE WHERE WE DON'T KNOW THE STRUCTURE OF THE ANTIBODY COMPLEXES. HERE WE SORT OF KNOW FROM THE STRUCTURE AND THIS CAN PROVIDE VALUABLE INFORMATION. SO WE ALSO UPLOADED OUR PROGRAM SO AGAIN, YOU ARE WELCOME TO DOWNLOAD AND FRY IT YOURSELF. -- TRY IT YOURSELF. SO A QUICK SUMMARY OF THIS SESSION, SO WE DEVELOPED THE ALGORITHM AIMED TO PREDICT ANTIBODY RESISTANCE FOR HIV-1 ANTIBODY AND WE'RE ABLE TO DEMONSTRATE THESE PREDICTORS TO DO THE WORK WITH HIGH ACCURACY. THE AIL -- ALGORITHM LOOKED AT CRITICAL RESIDUE TO UNDERSTAND THE ANTIBODY RESISTANCE. THIS PROGRAM IS AVAILABLE FOR DOWNLOAD TO TRY OUT. IN TERMS OF FUTURE DIRECTIONS, AS I MENTIONED EARLIER, WE'RE LOOKING AT DEEP LEARNING TO SEE IF FURTHER ACCURACY CAN BE IMPROVED. WE'RE APPLYING THE AIL GA -- ALGORITH ALGORITHMS FROM CENTERS TO HELP EXPLAIN THE DIFFERENCES TO OBSERVE TREATMENT ADVOCACY AMONG DIFFERENT PATIENTS BECAUSE IT'S INTERESTING TO SEE THE DATA FROM THOSE TRIALS. SOME ARE TREATED WITH THE SAME ANTIBODIES, SOME WORK WELL AND SOME DOESN'T. WE'RE TRYING TO USE THE PREDICTORS TO APPLY THEM TO THE ISOLATED VIRAL SCREEN TO SEE IF WE CAN DISCRIMINATE THE DIFFERENCE BETWEEN THE TREATMENT AND EFFICACY. SO WITH THAT, I WANT TO THANK ALL THE PEOPLE INVOLVED IN THE WORK I TALKED ABOUT TODAY AND TWO PEOPLE. DR. PETER QUANG AND THE POST-DOCTORIAL FELLOW WHAT HELPED DEVELOP THE ALGORITHM. WITH THAT I'D LIKE TO THANK YOU FOR YOUR ATTENTION. ARE THERE ANY QUESTIONS? I'M NOT IN EXPERT IN THE VIRUS. DID I WORK ON STABILIZING THE CONFORMATION BUT FOR THIS TYPE OF WORK LIKE WE DID FOR HIV, YOU HAVE TO HAVE A SET OF DIVERSE SEQUENCES. SO ONCE HAVE YOU A DIVERSE SEQUENCE YOU CAN HAVE DIFFERENT ANTIBODIES AND THEN YOU HAVE THE VARIABLE YOU CAN IDENTIFY WHICH POSITIONS ARE IMPORTANT. SO IN ORDER FOR THE VIRUS TO WORK, WE HAVE TO CHECK IF THERE'S A DIVERSITY WITHIN THE VIRUS. I KNOW IN THE SAME FAMILY THERE'S ALSO OTHER VIRUSES I FORGOT THE NAME BUT I THINK IT'S A DISTANCE SO YOU HAVE TO LOOK AT IT WITHIN THE VIRUS THERE'S ENOUGH VARIABILITY TO PERFORM THE SEQUENCE WITH. [COMMENT OFF MIC] >> WHEN WE LOOK AT THIS PROLINE. I JUST CHECKED AND THERE'S ANOTHER PROGRAM THAT CAME ON IN MAY AND IT'S MORE RELATED TO THE CRISPER. WE MIGHT LOOK AND IF THERE'S A PROBLEM WE WILL JUST CHANGE THE NAME. >> IN LOOKING AT THE SEQUENCES AND VALUE SO YOU CAN CLASSIFY IT ON YOUR OWN DEPENDING ON YOUR DESIGN. SORRY, EACH IS DESIGNED FOR ONE ANTIBODY. MAYBE I WASN'T CLEAR ABOUT THAT. YOU CAN DEVELOP THE PREDICTOR FOR THE ANTIBODY IN THE DATABASE THERE WILL BE 3 ENVELOPE SEQUENCES AND EACH ASSOCIATED WITH THE VALUE AGAINST THE ANTIBODY. SO YOU CAN CONVERT THIS BASED ON THE CODE. [COMMENT OFF MIC] >> THAT'S A SIMILAR QUESTION THE LADY OVER THERE HAD. IN ORDER FOR THIS TO WORK YOU NEED DIVERSITY AND FOR VIRUS SEARCHS I KNOW IT'S A CONSERVATIVE VIRUS AND THIS WOULD NOT WORK. SO YOU'RE ASKING WHERE TO RESPOND WITH THE DATA? YOU NEED TO HAVE -- YOU HAVE TO ALREADY HAVE THE DATA AND IN VACCINE DATA WE HAVE THE RESOURCE TO EVALUATE AN ANTIBODY OF INTEREST WITH -- WE HAVE A STANDARD OF 200 AND THERE ARE OTHER BIG HIV LABS AND 600 TO 200 ROUGHLY. YOU CAN ADD FEATURES AND YOU BASICALLY WANT TO ADD MORE RELATED INFORMATION IN THERE AS POSSIBLE. JUST ADD IT IN AND SEE WHAT HAPPENS. ANY PLORE -- MORE QUESTIONS? THANK YOU VERY MUCH.